Balyasny's $38B Playbook for Governing Frontier AI: Permissions, Not Just Model Choice

Balyasny Asset Management, a $38 billion asset manager, says Claude Fable 5 has cut its merger-arbitrage analysis time from three to five days down to roughly 30 minutes. The claim comes from an Anthropic case study describing how BAM runs Fable inside its proprietary BAMAgent platform, where the model independently plans a research task, selects tools, analyzes evidence, recovers from its own errors, and produces a finished deliverable. In BAM’s internal evaluation against thousands of real financial tasks, Fable hit an 89.4% success rate versus 86.1% for the firm’s prior production model, and solved internal economics problems that had gone unsolved before. Chief AI Officer Charlie Flanagan frames 2026 as the year AI shifted from “systems that do search” to “systems that do work.”

What’s more specific than the speed claim is how BAM governs it. Flanagan describes safety as “a product and operating-model question” rather than a model-selection question — meaning the controls that matter are approved data boundaries, least-privilege access, tool-level permissions, logging, mandatory human review, and defined escalation paths, deliberately built so a more capable model doesn’t automatically get more authority. BAM is now scaling past 300 agents running continuous analysis, describing the shift as moving from viewing AI as tools to viewing it as “teammates” that improve with repeated use.

BAM isn’t alone in making this bet at the research layer rather than back-office automation: T. Rowe Price expanded Claude across its investment organization the same week, with portfolio managers using it for research synthesis and developers building internal tools in Claude Code — governed by dedicated AI leaders inside each investment division under a central T. Rowe Price Labs group. President and CIO Eric Veiel put the value case simply: Claude lets investment professionals “cover more ground and go deeper on what matters.”

For consulting engagements advising on agentic AI governance, BAM’s model is a rare numbers-backed template: capability gains and expanded authority are treated as two separate decisions, verified against a real evaluation benchmark rather than a vendor’s claim.